TOPLINE

A combined model integrating mammographic AI, polygenic, and clinical risk scores significantly improved risk discrimination in breast cancer over 10 years compared with any individual model in a large observational study.

METHODOLOGY

• Breast cancer risk assessment still primarily relies on traditional clinical risk models. Recent studies have shown that adding polygenic risk scores improves risk discrimination, and that breast-tissue features captured by mammography AI can provide predictive information beyond traditional risk factors (including breast density). The value of combining all three risk domains has been unclear.
• To investigate, researchers conducted a prospective cohort study of 82,957 women enrolled in the Kaiser Permanente Research Bank between 2003 and 2020. All had a negative index screening mammogram and no prior breast cancer or high- or moderate-penetrance breast cancer gene mutation. About 21% were younger than 50 years and 75% were White.
• Mammographic AI risk scores (using Mirai deep-learning algorithm) were generated from the index screening mammograms; 5-year clinical risk scores (Breast Cancer Surveillance Consortium version 3) were based on patient surveys and electronic health records; and polygenic risk scores were based on 313 common gene variants (PRS313), adjusted for genetic ancestry.
• The models' performance in predicting breast cancer risk was evaluated alone and in combination with up to 10 years after the index screening mammogram.

TAKEAWAY

• During a median follow-up of 9 years, 2471 women developed breast cancer. T he combined AI, polygenic, and clinical model achieved the highest discrimination accuracy (C-index, 0.70) compared with AI-based model alone (0.66), the clinical model alone (0.62), or the polygenic risk score alone (0.61).
• In the combined model, the AI-based mammography risk model had the strongest association with breast cancer risk (hazard ratio [HR] per 1-standard deviation increase, 1.55), followed by the polygenic risk score (HR, 1.44) and the clinical risk model (HR, 1.28).
• Compared with the clinical model, the combined model identified more women who later developed breast cancer within the highest-risk decile, capturing 36% vs 27% of cancers during the first year and 26% vs 19% over 10 years.
• The combined model improved discrimination by 0.06-0.09 across racial and ethnic groups compared with the clinical risk model alone. Although performance did not differ significantly by race or ethnicity, it was slightly higher among Hispanic women and slightly lower among Asian and Black women vs White women.

IN PRACTICE

"This study demonstrates the complementary value of clinical, genetic, and mammography AI risk scores, highlighting the potential benefit of integrated models to enhance personalized decision-making for screening and prevention," the authors wrote.

SOURCE

The study, led by Vignesh A. Arasu, MD, PhD, Kaiser Permanente Northern California, Pleasanton, California, was published online in JNCI: Journal of the National Cancer Institute.

LIMITATIONS

Interpretation was limited by wide confidence intervals among minoritized populations, especially Black women. The clinical risk factor model was opportunistically evaluated on both invasive cancers and ductal carcinoma in situ, although it was trained to predict invasive cancers only. 

DISCLOSURES

The study was funded by the National Cancer Institute and The Permanente Medical Group Delivery Science and Applied Research. One co-author reported serving as an unpaid board member and receiving grant funding from the Quantum Leap Healthcare Collaborative for the I-Spy trial. Additional disclosures are noted in the original article.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.